Scene segmentation method of alternate update network based on multiple receptive fields

Abstract

The invention discloses a scene segmentation method of an alternate update network based on multiple receptive fields, which comprises the following steps: S1, performing feature map extraction on aninput image through a pre-trained convolutional neural network; S2, obtaining a feature map containing high-level semantic information through a pre-trained set network hole pyramid module; and S3, calculating classification loss of pixels one by one on the basis of the feature map obtained in the step S2 to obtain a segmentation heat map. According to the method, semantic segmentation of a sceneis carried out by using a cavity pyramid pooling network; on one hand, the integrated network cavity pyramid module can make full use of the feature map; the method improves the flow of information ina network, reduces the model parameters, achieves the compression of a model, increases the receiving field size of a convolution kernel through combining with an extended convolution method, achieves the segmentation of targets with different sizes in a scene graph, is high in robustness, is high in calculation efficiency, and the like.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a scene segmentation method based on an alternate update network with multiple receptive fields. Background technique [0002] Convolutional networks have been driving advances in various directions in the field of computer vision, such as object localization, object detection, and image recognition. At the same time, people also introduce convolutional networks into the problem of semantic segmentation, thereby replacing traditional manual labeling methods and achieving end-to-end segmentation output. The function of image recognition is to identify what is in an image, the result of target positioning is the position of the target, and semantic segmentation is to answer what categories are in the image and the specific positions of these categories from the perspective of pixel level. Semantic segmentation includes many sub-directions, such as: clothing analysis, scene segmentat...

AI Technical Summary

Problems solved by technology

[0004] (1) Feature resolution reduction: During the repeated pooling and downsampling process of the neural network, the feature resolution of the image will be reduced, resulting in the loss of some semantic information

[0005] (2) Too many network parameters increase the cost of training: With the development of convolutional neural networks, in order to obtain better segmentation results, people do not hesitate to increase the number of network layers to obtain better segmentation results, resulting in network parameters With the increase of the number of layers, the training difficulty of the network is increased, and the requirements for computing resources are increased.

[0006] (3) The size of the receptive field is limited: the scene graph includes a lot of scene objects, as small as a street lamp on the road, as large as a tall building, and their sizes vary greatly in the picture, so receptive fields of different sizes are required The convolution kernel can have a better segmentation effect

[0007] (4) The utilization rate of feature information is low: as the network deepens, higher-level semantic information is extracted, but the feature map information of the shallow network has not been fully utilized

Images